关于train.py中的的Dataset问题

[zhibeiyou135]

在train.py中：
if args.dataset == 'mvtec':
train_data = MVTecDataset(root=args.train_data_path, transform=preprocess, target_transform=transform,
aug_rate=args.aug_rate)
并未设置mode='train',而 MVTecDataset中mode的默认值为 'test' 。下述代码：
if mode == 'train':
# 如果模式是 'train'
self.cls_names = [obj_name] # 将对象名称添加到类别名称列表
save_dir = os.path.join(save_dir, 'k_shot.txt') # 构建保存目录的路径
else:
self.cls_names = list(meta_info.keys()) # 否则，获取所有类别名称
无论是test.py还是train.py都不会执行 mode == 'train'。请问是应该将mode默认值设置为train，还是我理解的幼体，在train.py中mode就应该为test？

[ByChelsea]

Hi, in the zero-shot setting, we use the test set of another dataset for training (this conforms to the principle that zero-shot learning cannot train on data of related objects). More details please refer to our technical report.

The train mode in dataset.py is written for the few-shot setting.

[siri-ius]

Hi,

I have four following questions based on your answer:

> The train mode in dataset.py is written for the few-shot setting.

1. Does it mean that when we train VAND model using the command 'bash train.sh', it is being trained only as a few-shot learning model?
2. If yes, and then we store checkpoints for VisA and MVTec-AD datasets in exps/visa/new_checkpoints/epoch_15.pth and exps/mvtec/new_checkpoints/epoch_3.pth. Should we use these checkpoints to evaluate the perfomance of VAND on mvtec and visa datasets as in ZSAD (bash test_zero_shot.sh) or FSAD (bash test_few_shot.sh) ?
3. If answer to Q1 is yes, then could you please guide us on how can we train VAND as a zero-shot learning model?
4. If answer to Q2 is FSAD, then could you please guide us on how can we evaluate VAND as a zero-shot learning model?

Thanks a lot!
Looking forward to hearing from you!